Physical Execution Plans for Model-Driven Engineering: Decoupling DSL Design from Runtime-Specific Representations
TL;DR
This work argues that physical execution should be a first-class modeling concern and proposes physical execution plans for Model-Driven Engineering, which complements meta-models with execution plans that declare layouts, relations, indexes, access paths, and target runtimes without changing model meaning or modeling tasks.
Abstract
Model-Driven Engineering (MDE) raises the level of abstraction by representing systems using domain-specific languages (DSLs). However, the environment in which a DSL is developed frequently dictates how its artifacts are physically represented, accessed, and executed. Choices such as memory layouts, indexes, and runtime targets are typically inherited from the modeling framework rather than explicitly modeled. For instance, in the Eclipse Modeling Framework (EMF), models become object graphs in the Java Virtual Machine, limiting the control of DSL designers. We argue that physical execution should be a first-class modeling concern and propose physical execution plans for Model-Driven Engineering. Our approach complements meta-models with execution plans that declare layouts, relations, indexes, access paths, and target runtimes without changing model meaning or modeling tasks. Inspired by user-schedulable languages and progressive lowering, our approach separates model meaning from execution. We evaluate our approach over a set of modeling tasks, generating plan-specific Java and Rust backends that we compare against a conventional EMF baseline. These executors achieve task-dependent speedups of up to five orders of magnitude, while producing identical results. The measurements expose trade-offs among plans, opening a research agenda on cost-driven plan selection.